High-efficiency Screening Of Zeolite Materials For Adsorption Separation CO₂ By Machine Learning

The problem of global warming is becoming more and more serious.In order to reduce CO₂ in the atmosphere,researchers develop continuously cost-effective and energy-efficient CO₂ capture processes.Pressure Swing Adsorption（PSA）-based on CO₂ capture has emerged as a promising approach,which is widely used in gas separation backgrounds such as CO₂/N₂separation in post-combustion flue gas and CH₄/CO₂ separation in natural gas purification.However,the gas separation process based on adsorption requires the support of high-performance adsorption materials.Zeolite is widely used in this field because of its unique framework structure and pore feature.When zeolite adsorption materials are screened,single component adsorption isotherms for the target gas are typically the only information that can be reasonably obtained.At present,the determination methods of adsorption isotherms are mainly experimental method and molecular simulation method.which consume time and money.With the rapid development of computer technology,machine learning（ML）has attracted more and more attention.ML can model Structure-Adsorption Relationship（SAR）of zeolites based on existing data,and predict the adsorption isotherms more conveniently and quickly.As a result,ML method combined with appropriate screening strategy can be used to save time and efficiently find and design zeolite adsorption materials with excellent performance.Main contents of this thesis are concluded as following:（1）A structure descriptor"natural building unit（NBU）"of zeolite materials is associated with adsorption properties through ML.When using the ANN model,due to the explosion of its hyperparameter combination,a greedy parameter tuning strategy is used to improve the efficacy of model training and testing.The strategy can also avoid overfitting of the model.Finally,the performances of ANN models are verified to ensure that models could be used in the screening of zeolite adsorption materials for CO₂/N₂ separation in post-combustion flue gas and CH₄/CO₂ separation in natural gas purification.（2）A machine learning（ML）-based screening framework for zeolite adsorption materials is established.Firstly,the NBU descriptors of zeolites are calculated,and then the pure component adsorption data of the zeolites are predicted by ML model.Next,the mixture adsorption data that can accurately characterize the gas separation problem are predicted by ideal adsorbed solution theory（IAST）.Finally,candidate zeolite materials are selected by some adsorbent evaluation metrics.In this paper,taking CO₂/N₂ separation as an example,11 and 7candidate zeolites satisfying all the constraints of adsorbent metrics are screened from the IZA-SC database（194 zeolite materials）and the hypothetical zeolite database（504 zeolite materials）respectively,which demonstrates the practical value of the framework.（3）A screening framework for zeolite adsorption materials based on machine learning（ML）and idealized PSA is proposed.Based on the framework in Chapter 3,the performance indicators of idealized PSA process model are added to further screen zeolite materials.The reason is that the adsorbent evaluation metrics can only screen materials from the properties of the materials themselves,and cannot accurately predict the applicability of adsorption materials in PSA process.The screening of zeolite materials based on the idealized PSA process model includes using IAST to predict the gas adsorption data of the mixture,preliminary screening of zeolite materials according to the adsorbent evaluation metrics,and selection of the best candidate zeolite based on the performance indicators of the PSA process model.The framework is applied to the separation of CH₄/CO₂ under 4 groups of molar compositions（20%,30%,40%,50%CO₂）.The best zeolite materials meeting the requirements are successfully screened from IZA-SC database and hypothetical zeolite database,which proved the effectiveness of the framework.